Among-ASCII
What if LLM-based agents played Among Us in a traditional roguelike format?
- Python
Repository: https://github.com/albrescamily/among-ascii
Since Among us is by FAR my favorite game, I thought “What if I make LLMs play Among Us against each other?”. Its a perfect environment for LLM-based agents!
Its a social deductive game that we have to act and deduce who the impostor is based on observations and partial information.
But I faced I’m issue on this idea, How would I clone the Among Us game to make agents play against each other locally and evaluate their behaviors?
I can’t copy the original game by myself since this is not on the scope of my project, and I tried to use replicas I found on GitHub made by other users, but the messy code made the learning curve too high to keep going on this approach.
To give a context, after some weeks trying to adventure myself into traditional roguelikes games, and questioning myself If I was too dumb to play them (I honestly gave up on cogmind and Dungeon Crawl), I started researching about RL and game gymnasiums.
After I found out about NetHack Learning Environment, I came with a brilliant idea!
What If I ask ChatGPT Astra to recreate the Skeld map in ASCII?
Since the scope of this project is to build the agents from scratch, I decided not to spend too much time manually creating the environment and used AI to help generate its initial structure.
The environment is designed to simulate the traditional Among Us map and its main game mechanics. I added vents and sabotage systems to support impostor behavior, as well as tasks that crewmates can complete throughout the game.
The main goal is not to perfectly recreate Among Us, but to provide a sufficiently rich environment for experimenting with agent behavior, interaction, decision-making, and coordination.